https://gcc.gnu.org/bugzilla/show_bug.cgi?id=77436
Richard Biener <rguenth at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Keywords| |wrong-code
Component|c |middle-end
Known to work| |4.9.4
Target Milestone|--- |5.5
Summary|Incorrect constant result |[5/6/7 Regression]
|for summing loop inserted |Incorrect constant result
| |for summing loop inserted
--- Comment #4 from Richard Biener <rguenth at gcc dot gnu.org> ---
Confirmed, caused by SCCP (or SCEV/niter analysis):
Analyzing # of iterations of loop 1
exit condition [-2147483645, + , 1] != 2147483647
bounds on difference of bases: 4294967292 ... 4294967292
result:
# of iterations 4294967292, bounded by 4294967292
Replacing uses of: sum_11 with: 2147483648
while it computes correctly the evolution of the sum IV:
{0, +, {2147483650, +, 1}_1}_1
somehow the exit PHI value is computed as 2147483648 by
compute_overall_effect_of_inner_loop on that IV + x
{2147483650, +, {2147483651, +, 1}_1}_1
number of latch executions is computed as 4294967292.
We run into chrec_evaluate with the above where binominal coeff evaluation
checks for overflow in the signed type but the computation in chrec_evaluate
doing
arg0 = fold_build2 (MULT_EXPR, type,
CHREC_LEFT (chrec), binomial_n_k);
has no such checks. On the GCC 4.9 branch we fail:
(chrec_apply
(varying_loop = 1
)
(chrec = {2147483650, +, {2147483651, +, 1}_1}_1)
(x = 4294967292)
(res = scev_not_known))
not replacing:
sum_2 = PHI <sum_4(3)>
thus this is a regression with the wide-int conversion somehow.
